Semantic memory and psychological semantics

After setting the problem of semantic memory in psycholinguistic perspective, we began our account of how semantic propositions are verified. We first assumed a semantic feature representation which distinguished between defining and characteristic features. This representation was then coupled with a two-stage processing model, and we applied the resulting Feature Comparison model to the results of studies requiring the verification of simple subset statements. This model offered an explicit explanation of semantic relatedness and category size effects in this paradigm, while our assumption about characteristic features further suggested some new insights about the representation of natural language concepts and propositions. The Feature Comparison model was then extended to accommodate findings from recent Same-Different experiments. Here, the extended model proved capable of encompassing a range of semantic relatedness findings, including some newly reported effects which seem problematic for other models. We next applied the basic Feature Comparison model to the verification of simple property statements. While the representation of property information necessitated several new structural and processing considerations, the basic model again provided an explanation of many of the semantic effects on verification. Further, it suggested alternative interpretations of findings originally attributed to nonsemantic factors. We then considered the model in relation to explicitly quantified statements, and extended our processing assumptions so as to include several different natural language quantifiers. Once more the Feature Comparison model, when suitably augmented, was consistent with the existant data, including some new results which highlighted the decision component of the model. In the last section, we explored the utility of our semantic feature representation for understanding the nature of induction, analogy, and metaphor. Here, the processing routines differed from our usual rwo-stage model, and these new procedures offered a promising account of our experimental results on induction and analogy. Thus, we have been able to provide a systematic account of a diverse range of semantic phenomena in terms of a fixed representational scheme and a limited set of processes that compare feature sets. But although we consider these results very promising, we note that there are several gaps. First among these is the fact that we have not offered a sufficiently detailed account of the second stage of the Feature Comparison model. This was particularly noticeable in our treatment of quantified statements (Section V), where our lack of detail about the second stage precluded us from drawing any strong predictions from our range-of-features hypothesis about quantifiers. Secondly, although we provided a detailed account of property statements containing is and has constructions (Section IV), we did not deal with those statements that involve more complex verbs. Third, our treatment of induction and analogy (Section VI) was limited to a restricted set of problems and data, while our handling of metaphor was confined to but a few remarks. Lastly, throughout this chapter we have emphasized the verification of single propositions and have not concerned ourselves with the semantic integration problems posed by multiproposition statements. While this is a formidable list of omissions, all of them seem, in principle, to be amenable to a semantic features approach. They therefore constitute topics for future work, rather than indications that our theoretical orientation is misdirected. Indeed, we consider the semantic feature direction taken here to be the most profitable one for understanding the domain called semantic memory, and for relating it to other domains of natural language processing.